Semantic segmentation of point clouds is one of the main steps in automated processing of data from Airborne Laser Scanning (ALS). Established methods usually require expensive calculation of handcrafted, point-wise features. In contrast, Convolutional Neural Networks (CNNs) have been established as powerful classifiers, which at the same time also learn a set of features by themselves. However, their application to ALS data is not trivial. Pure 3D CNNs require a lot of memory and computing time, therefore most related approaches project ALS point clouds into two-dimensional images. Sparse Submanifold Convolutional Networks (SSCNs) address this issue by exploiting the sparsity often inherent in 3D data. In this work, we propose the application of SSCNs for efficient semantic segmentation of voxelized ALS point clouds in an end-to-end encoder-decoder architecture. We evaluate this method on the ISPRS Vaihingen 3D Semantic Labeling benchmark and achieve state-of-the-art 85.0% overall accuracy. Furthermore, we demonstrate its capabilities regarding large-scale ALS data by classifying a 2.5 km2 subset containing 41 M points from the Actueel Hoogtebestand Nederland (AHN3) with 95% overall accuracy in just 48 s inference time or with 96% in 108 s.
Since trees are a vital part of urban green infrastructure, automatic mapping of individual urban trees is becoming increasingly important for city management and planning. Although deep-learning-based object detection networks are the state-of-the-art in computer vision, their adaptation to individual tree detection in urban areas has scarcely been studied. Some existing works have employed 2D object detection networks for this purpose. However, these have used three-dimensional information only in the form of projected feature maps. In contrast, we exploited the full 3D potential of airborne laser scanning (ALS) point clouds by using a 3D neural network for individual tree detection. Specifically, a sparse convolutional network was used for 3D feature extraction, feeding both semantic segmentation and circular object detection outputs, which were combined for further increased accuracy. We demonstrate the capability of our approach on an urban topographic ALS point cloud with 10,864 hand-labeled ground truth trees. Our method achieved an average precision of 83% regarding the common 0.5 intersection over union criterion. 85% percent of the stems were found correctly with a precision of 88%, while tree area was covered by the individual tree detections with an F1 accuracy of 92%. Thereby, we outperformed traditional delineation baselines and recent detection networks.
27 March 2020 in PFG – Journal of Photogrammetry, Remote Sensing and Geoinformation Science
While most research in automatic semantic segmentation of 3D geospatial point clouds is concerned with enhancing respective Machine Learning (ML) models, we aim to shift the focus to be more of a data-centric nature. This means, we consider the creation of respective data sets that ML models learn from as key component, since even the most sophisticated model performs poorly when learning from suboptimal data. In this regard, the straightforward approach of providing labeled data abundantly is prohibitively expensive and just not scalable in times of high-frequency data acquistion cycles, where a dedicated training set should be available for each new epoch, as ML models often lack generalizability. As a remedy, we rely on Active Learning (AL), which is a cost-efficient and quick method to generate required training data at scale. Although AL has been (scarcely) applied in the geospatial domain before, a comprehensive evaluation of its capabilities, including benchmarking of achievable accuracies is lacking. Therefore, we apply the AL concept to both ISPRS’ current point cloud benchmark data sets as well as to a third large scale National Mapping Agency point cloud. Respective experiments are conducted with both a feature-driven Random Forest classifcation approach and a data-driven Submanifold Sparse Convolutional Neural Network classifier. Our experiments verify that by labeling only a fraction of available training points (typically ⪡ 1%), we can still reach accuracies that are at maximum only about 5 percentage points worse compared to leading benchmark contributions.